Korean family benefits, in one place
Korean parents are entitled to support from national, provincial, city and district programmes at the same time. The amounts change with where you live, which child this is, and how long you have lived there. Most families never see the full picture. 혜택모아 assembles it — with the government document behind every figure.
보자기 — separate pieces, one cloth
The name
The name is the product description. Nothing here is hidden from families — every one of these programmes is published by the government that runs it. They are simply published in dozens of different places, in different formats, under rules that change by district and by year.
The problem
These are real municipal birth-grant structures. A family that moves one district over, or has a second child rather than a first, is in an entirely different scheme — and there is no single place that tells them so.
And the danger runs both ways. A headline figure of ₩19,000,000 is not a cheque — it is ₩300,000 a month for sixty months, and only if the family stays registered in the county. Showing that number without its conditions is not generosity. It is a promise the government never made.
What it does
A short first pass covers most programmes. Extra detail is requested only where it genuinely changes an amount, not to fill a form.
Instalments are shown as instalments. Residence requirements are shown as requirements. A five-year total is never presented as cash on the day.
Each benefit shows the government page or ordinance the number came from, so a family — or an official — can check it.
Wireframes — build screens are being captured and will replace these. The coloured rule on the third screen is the citation line.
How the numbers are kept honest
Build journal
Development runs in numbered working sessions, each opening by verifying the previous session's record is intact and closing with a written, committed handoff. These are the moments that shaped the product.
The first real fork: a benefit whose value genuinely varies could be flattened to a single ceiling figure with a disclaimer, or the variation could be modelled properly in the schema. Flattening was faster and would have been wrong on screen. The principle that came out of it governs everything since.
The obvious build hands a family's details to a language model and lets it answer. It demos beautifully and it is indefensible — you cannot explain to a government partner why the number was what it was. So money and words were put on separate rails, and the model was locked out of the arithmetic entirely.
Asking everything up front is how benefit finders lose people. Most programmes resolve from a handful of facts; only some need more. The matcher was built to answer what it can from stage one and clearly defer the rest, rather than demanding a full profile before showing anything.
A long stretch of work spent not building features at all: reading municipal pages and ordinances, one benefit at a time, and recording what each value actually rests on. It produced a deliberately narrow vocabulary — separating "we have not checked this" from "we checked, and the source is wrong," which most review processes collapse into one word.
A district grant stored four payment tiers and two instalment schedules. The district's own page returned all six figures, identical, down to the contact phone number. A clean pass by any normal standard.
Except the page carried a last-modified date of 2022 and described a national voucher under a rule that had since changed. The value was perfectly faithful to a source that had stopped being true — so the vocabulary gained a verdict for exactly that, and the entry moved off a status it did not deserve.
A sweep across the whole benefit corpus looked for cross-references pointing at programmes that did not exist. It returned nothing. Clean.
It was not. A scoping error made every comparison ask whether a list contained itself — always true, so nothing was ever flagged. A broken filter and a healthy corpus produce identical output. The only thing that caught it was a rule adopted earlier: before running a check, write down one case that must appear. One had been written down. It did not appear.
Underneath was the real finding — a set of rules marking benefits a household can only receive one of, declared in the data and applied by no code at all. Every one of them would have inflated a family's total.
Each session logs its own errors with the same prominence as its results — a prediction made without measuring, a search that matched a false positive, a guard written and then wired to nothing. A record that contains only successes tells a reader nothing about whether to believe it.
A retraction is cheaper than a wrong record — and a check stated in advance is cheaper than both.
Where it stands
There is no launch date announced, and there will not be one until the verification work is further along. An app that tells families what money they are owed does not get to ship a number it has not checked.
Colophon
혜택모아 is designed and built by Vincent Kim, in Richmond, British Columbia. The approach comes from a previous life running contact-centre workforce management — queue routing, forecasting, and the habit of not shipping a diagnosis you cannot defend.
The development record is kept as a written audit trail rather than a changelog, because the intended standard is a government partnership, and a partnership like that asks a question a screenshot cannot answer: where did this number come from?